Enhancing untargeted metabolite elucidation by machine learning of fragmentation in mass spectrometry

About This Grant

PROJECT ABSTRACT Metabolites are essential modulators, biomarkers, and signaling molecules in human health and disease, yet most metabolomic data remain unannotated due to limitations in computational workflows for liquid chromatography tandem mass spectrometry (LC-MS/MS). Nearly 90% of MS/MS features in untargeted metabolomics studies lack structural assignments, hindering mechanistic and translational discovery. Deep learning-based in silico fragmentation has improved annotation accuracy, but existing models depend on proprietary data, lack confidence measures, and are restricted to one-at-a-time structure prediction. This K99/R00 project will develop open, robust, and scalable workflows for metabolite structural elucidation. During the mentored K99 phase, the candidate will expand the ICEBERG geometric deep learning model to train entirely on curated open-source datasets (GNPS, MassSpecGym) through knowledge distillation from proprietary models and introduce atom-level confidence scoring analogous to pLDDT in protein folding. The open model will provide interpretable confidence maps and enable high-confidence substructure annotation without commercial data dependence. During the independent R00 phase, network-based reasoning will be incorporated to jointly analyze chemically related spectra through integer-linear optimization and graph-based propagation. This framework will create an open metabolite atlas by repository-wide substructure annotation of Pan-ReDU (with spectra and metadata curated from GNPS, Metabolomics Workbench, etc), empowering large-scale reanalysis and hypothesis generation. Proof-of-principle studies in cancer metabolism, inflammatory bowel disease and mitochondrial disease cohorts will demonstrate the biological relevance and translational potential of the approach. The candidate’s long-term goal is to establish an independent research program at the intersection of artificial intelligence and metabolomics, focusing on comprehensive elucidation of metabolites that drive biology, disease, and therapeutic discovery. The career development plan includes training in computational chemistry, untargeted metabolomics, and clinically relevant disease biology; mentorship from leading experts at MIT, Harvard, and the Broad Institute; and structured professional development in grant writing, teaching, and leadership. The institutional environment at MIT and its aZiliates provides exceptional computational, experimental, and translational resources, including access to high-performance computing, state-of-the-art LC-MS/MS facilities, and large clinical metabolomics datasets. Together, these resources and mentorship will ensure the successful transition to research independence and leadership in AI-driven metabolomics.